A set of scripts and tools to benchmark popular ML models with IREE, optimized with compiler development workflows in mind. The goal is to prioritize the ease of set up over completeness or production readiness, while providing a good enough proxy for real-world deployment.
All the models below were exported using SHARK AI.
The currently supported models are:
- SDXL (fp16, int8, fp8)
- Llama3 8b
The MLIR files checked in are using Git LFS. To checkout the files locally
- Make sure you have git-lfs installed locally. On Linux that is
sudo apt-get install git-lfs
- Install Git LFS for your user account
git lfs install
- To checkout the MLIR files locally
git lfs pull